Inferring Quantitative Trait Loci Using a Bayesian Variable Selection Model and Markov Chain Monte Carlo Convergence Diagnostics

• Activity Number: 76
• Type: Contributed
• Date/Time: Sunday, August 6, 2006 : 8:00 PM to 9:50 PM
• Sponsor: Biometrics Section
• Author(s): Daniel Shriner*+ and Nengjun Yi
• Companies: The University of Alabama at Birmingham and The University of Alabama at Birmingham
• Keywords: convergence diagnostics ; variable selection ; quantitative trait loci

Abstract

We previously developed a Bayesian composite space model for the genome-wide identification of quantitative trait loci. Since the number of quantitative trait loci is unknown, we use a variable selection method to address the unknown dimensionality of the model. In this work, we examine the behavior of our Markov chain Monte Carlo algorithm using standard convergence diagnostics. The degrees of cross-correlation and autocorrelation for the primary variables in the model were not very high. There was no significant burn-in period for 93% of the runs. The number of loci with inclusion probabilities reaching significance as assessed by Bayes factors was highly insensitive to the prior number of included loci, although effect sizes were easily overestimated. We suggest three possible criteria for determining the most parsimonious subset of variables for which the data support inclusion.